The accuracy of the maps (random forests part of the model) was assessed using 5-fold cross-validation, with model re-fitting, and reported using RMSE and % variance explained. The latter is defined as 1-MSE/sigma-square, where MSE is the mean square error at cross-validation points and sigma-square is the variance of the target variable. Some soil properties were log-transformed prior to prediction using regression kriging. In such cases the maps are the back-transformed regression kriging maps, but the % variance explained was derived in the transformed (trans-Gaussian) space which differ from the % variance explained in the original space. The figure and table below shows scatter plots and summary statistics of cross-validation results.
Example of a mapping accuracy assessment plot for soil organic carbon, pH and bulk density (obtained with 5-fold cross-validation).
Table: Mapping accuracy for AfSoilGrids250m maps assessed using 5-fold cross-validation (using only random forests models).